[roji]

Fixes #3566.

json_to_recordset/jsonb_to_recordset do not decode the base64 representation used for byte[] properties inside JSON. Collection expansion therefore returned the base64 text's bytes, even though full-document materialization and indexed property access returned the original payload.

Related upstream fix: dotnet/efcore#33440, which fixed dotnet/efcore#33435. That addressed the same base64-versus-binary-conversion mismatch in SQL Server's scalar JSON access by rewriting JSON_VALUE to OPENJSON ... WITH. This PR addresses PostgreSQL's separate collection-recordset path; Npgsql's scalar JSON decoding already works. The late JSON postprocessing approach follows the same upstream precedent while retaining flat PostgreSQL SQL.

Extract affected recordset columns as text and wrap their references in decode(..., 'base64'). This is applied after structural projection expansion, which requires column expressions, and avoids adding SQL subqueries. Existing scalar JSON decoding is unchanged.

Add sync/async regression coverage using complex JSON collection mapping in the shared JSON translation fixture for both json and jsonb, comparing full materialization with projected bytes and checking server-side binary equality. Include non-ASCII bytes, empty arrays, and null values.

Confirmed with a standalone console repro on released providers 8.0.0, 9.0.4, 10.0.0, 10.0.3, and unmodified main. The equivalent collection projection works with SQL Server 10.0.0. The original repro returns matching bytes after the fix. Manual checks also confirmed owned and complex JSON mapping.

Validation: dotnet test --project test/EFCore.PG.FunctionalTests --filter-class '*Json*' — 1,757 passed, 9 skipped, using an isolated PostgreSQL Testcontainer. After switching the regression to complex mapping, the affected JsonTranslationsNpgsqlTest suite passed all 12 cases.

This is a bug fix targeting main (11.0), not a new JSON feature or a serialization-format change. No backports are included.

[Copilot]

🟢 Approval recommended

The targeted rewrite and regression coverage address the reported inconsistency, with no blocking issues identified.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

The focused rewrite and regression coverage address the reported mismatch with no unresolved blocking findings; tests were not independently rerun.

No open issues 🎉

_{Review effort: Balanced}